Multi-agent LLM systems are built by fixing a cast of roles and then designing how the cast coordinates. Which agents should make up the group for a given task is left to the designer, and the same cast is reused for every task. Financial trading makes the cost of that choice visible: stocks trade differently, and the trading styles suited to one need not suit another. We study task-conditioned persona selection, composing the group from the characteristics of the task it will face, and instantiate it with FinPanel, an LLM multi-agent trading system in which a recruiter selects a panel of named trader personas for each stock from a direction-blind statistical sketch of how the stock trades, each persona gives an independent daily call in its own style, and a manager commits the trade after reading the panel's arguments and their desk-kept, hierarchical track records. Because nothing is returned to the traders, the panel stays the panel that was selected, and it can trade a new stock from an empty record. On eight US stocks over the out-of-knowledge year 2025, against three prior LLM trading agents re-implemented in the same framework and against rule-based, predictive and reinforcement-learning baselines, the selected panel is ahead of a fixed panel of the same size and of the prior agents run as their papers prescribe, with a shallower drawdown than buy-and-hold on every stock; it keeps its performance with no warm-up and with no record at all, whereas the memory-augmented agents' returns swing with their warm-up; and neither more seats nor the manager's reading of the arguments substitutes for selecting the right ones. Who is in the group is a design decision to be made per task, before how the group coordinates.
Figures & tables
Figure 1: Overview of FinPanel. The recruiter reads a direction-blind sketch of the stock and seats N named traders once per run; every day each trader reads the same brief and gives an independent call with an argument; the distiller settles each call against the market and maintains a hierarchical track record per trader; the manager reads the arguments and the records and commits the portfolio. The six names shown are the panel recruited for TSLA.
Figure 2: Cumulative return over 2025-FY per stock for FinPanel, the three LLM agents run as their papers prescribe, and buy-and-hold.
FinMem/FinAgent stocks
Added stocks
Mean over the eight stocks
Category
Method
TSLA
NFLX
AMZN
MSFT
GS
WMT
BA
XOM
CR% ↑
SR ↑
SoR ↑
MDD% ↑
Market
Buy-and-hold
19.2
5.0
4.5
16.7
55.8
25.4
27.4
16.2
21.3
0.68
0.96
−28.3
Rule-based
SMA cross
16.1
8.8
−7.5
13.8
34.4
−3.2
1.7
6.4
8.8
0.34
0.48
−21.0
WMA cross
26.8
−4.5
−9.2
5.0
41.4
1.3
17.4
−8.6
8.7
0.25
0.39
−18.5
Bollinger
5.3
29.4
9.2
10.2
6.2
4.7
35.9
13.3
14.3
0.63
1.16
−15.2
Trend following
6.2
6.1
12.4
16.0
47.9
10.1
−9.7
−28.3
7.6
0.16
0.38
−15.6
Table 1: Performance comparison on 2025-FY. Cumulative return (%) per stock and the mean CR, Sharpe ratio, Sortino ratio and maximum drawdown over the eight stocks (MDD: higher is shallower). Every method is in its published configuration, and Table 4 reports the LLM agents without their warm-up; all LLM methods run on gpt-oss-120b with the same inputs and costs. Among the LLM methods, the best per column is in bold and the second best underlined; the FinPanel row is shaded.
FinMem/FinAgent stocks
Added stocks
Mean over the eight stocks
Category
Method
TSLA
NFLX
AMZN
MSFT
GS
WMT
BA
XOM
CR% ↑
SR ↑
SoR ↑
MDD% ↑
Market
Buy-and-hold
−15.2
48.5
0.4
18.3
20.2
8.5
25.1
3.1
13.6
0.85
1.20
−26.9
Rule-based
SMA cross
−10.8
28.9
−0.6
20.9
3.9
−6.8
−0.7
−1.1
4.2
0.54
1.11
−18.9
WMA cross
15.4
11.3
0.4
4.7
15.4
−1.3
10.5
−13.0
5.4
0.38
0.53
−16.7
Bollinger
−4.0
36.5
1.6
11.8
4.7
4.7
33.6
13.3
12.8
1.07
2.48
−13.4
Trend following
3.3
22.5
14.1
19.2
29.3
8.4
−4.0
−16.6
9.5
0.90
1.58
−11.1
Table 2: Performance comparison on 2025-H1 (the April 2025 tariff shock and the recovery that followed); columns as in Table 1 , with the Sharpe and Sortino ratios annualised from the half-year.
2025-H1
2025-FY
Panel
CR ↑
SR ↑
MDD ↑
CR ↑
SR ↑
MDD ↑
FinPanel
11.7
0.86
−18.2
16.1
0.60
−20.7
w/o selection
2.6
0.13
−22.9
7.2
0.24
−25.1
w/o sketch
4.7
0.24
−21.5
7.9
0.28
−26.1
Table 3: Effect of persona selection: mean cumulative return (%), Sharpe ratio and maximum drawdown (%) over the eight stocks on 2025-H1 and 2025-FY for the recruited panel (FinPanel), the fixed panel seated on every stock (w/o selection), and the panel recruited from the ticker alone (w/o sketch). Best per column in bold; per-stock values are in Appendix C.1 .
2025-H1
2025-FY
Method
CR ↑
SR ↑
MDD ↑
CR ↑
SR ↑
MDD ↑
FinPanel, 2-y
11.7
0.86
−18.2
16.1
0.60
−20.7
FinPanel, zero
11.5
0.86
−17.2
15.6
0.61
−19.5
FinPanel, no record
13.1
0.87
−16.3
16.2
0.55
−20.1
FinMem, 2-y
1.9
0.15
−21.6
7.9
0.15
−23.2
FinMem, zero
10.5
0.70
−23.8
18.6
0.60
−25.3
Table 4: Effect of the warm-up: mean cumulative return (%), Sharpe ratio and maximum drawdown (%) over the eight stocks on 2025-H1 and 2025-FY for each LLM method with the two-year warm-up its paper prescribes (2-y) and with none (zero), and for FinPanel also with the record removed (no record). Best per column in bold and second best underlined; the FinPanel rows are shaded; TradingAgents is memoryless. Per-stock values are in Appendix C.2 .
Figure 3: Mean cumulative return and maximum drawdown against panel size N∈{1,4,6,8} on 2025-H1 and 2025-FY, with recruited panels and the two-year record; the light markers are the eight per-stock values. Per-stock values are in Appendix C.3 .
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
FinMem/FinAgent stocks
Added stocks
Mean over the eight stocks
Window
Method
TSLA
NFLX
AMZN
MSFT
GS
WMT
BA
XOM
CR% ↑
SR ↑
SoR ↑
MDD% ↑
FinPanel
5.5
52.8
−0.1
10.2
5.4
8.0
10.8
1.1
11.7
0.86
1.34
−18.2
w/o selection
−16.9
39.7
−15.4
7.5
8.0
−4.9
1.1
1.3
2.6
0.13
0.16
−22.9
2025-H1
w/o sketch
−9.4
52.7
−10.4
3.3
−0.6
9.5
0.3
−7.4
4.7
0.24
0.44
−21.5
FinPanel
14.5
16.1
5.3
4.9
37.2
25.0
12.5
13.3
16.1
0.60
0.82
−20.7
w/o selection
−17.0
8.5
−10.4
4.3
44.9
10.3
−1.9
18.7
7.2
0.24
0.29
−25.1
Appendix
Table 5: Effect of persona selection per stock: cumulative return (%) per stock and the mean CR, Sharpe ratio, Sortino ratio and maximum drawdown over the eight stocks, on 2025-H1 and 2025-FY, for the rows of Table 3 .
FinMem/FinAgent stocks
Added stocks
Mean over the eight stocks
Window
Method
TSLA
NFLX
AMZN
MSFT
GS
WMT
BA
XOM
CR% ↑
SR ↑
SoR ↑
MDD% ↑
FinPanel, 2-y
5.5
52.8
−0.1
10.2
5.4
8.0
10.8
1.1
11.7
0.86
1.34
−18.2
FinPanel, zero
18.6
48.0
−4.9
8.5
3.7
8.5
7.2
2.3
11.5
0.86
1.50
−17.2
FinPanel, no record
9.3
54.0
−2.3
14.9
11.6
6.7
3.8
6.5
13.1
0.87
1.30
−16.3
FinMem, 2-y
−33.1
30.6
−6.1
5.5
19.3
12.1
0.0
−12.7
1.9
0.15
0.22
−21.6
FinMem, zero
−9.8
48.3
−0.1
7.1
19.3
7.7
12.9
−1.5
10.5
0.70
0.94
−23.8
Appendix
Table 6: Effect of the warm-up per stock: columns as in Table 5 , for the rows of Table 4 .
FinMem/FinAgent stocks
Added stocks
Mean over the eight stocks
Window
Method
TSLA
NFLX
AMZN
MSFT
GS
WMT
BA
XOM
CR% ↑
SR ↑
SoR ↑
MDD% ↑
N=1
−19.1
64.7
−18.0
14.7
−1.8
9.5
20.7
−8.1
7.8
0.48
1.12
−20.8
N=4
−7.6
40.7
−3.6
2.8
12.1
3.2
−4.6
−9.7
4.2
0.12
0.22
−19.0
N=6
5.5
52.8
−0.1
10.2
5.4
8.0
10.8
1.1
11.7
0.86
1.34
−18.2
2025-H1
N=8
−13.0
39.8
−9.8
5.7
1.0
1.9
3.7
6.8
4.5
0.29
0.41
−21.3
N=1
−19.7
33.9
−14.7
13.0
26.4
25.7
13.6
−2.6
9.5
0.29
0.42
−24.1
Appendix
Table 7: Panel size per stock: columns as in Table 5 , at N∈{1,4,6,8} with recruited panels and the two-year record, behind Figure 3 .
FinMem/FinAgent stocks
Added stocks
Window
Method
TSLA
NFLX
AMZN
MSFT
GS
WMT
BA
XOM
Mean
Buy-and-hold
−0.11
2.33
0.13
1.22
1.12
0.63
1.21
0.27
0.85
FinMem, 2-y
−1.64
1.79
−0.44
0.53
1.08
0.97
0.00
−1.12
0.15
FinMem, zero
0.03
2.59
−0.02
0.59
1.08
0.68
0.73
−0.09
0.70
FinAgent, 2-y
−1.42
1.64
−2.77
1.32
−0.32
−0.00
2.99
−1.64
−0.02
FinAgent, zero
−2.99
1.68
−1.60
0.70
−0.06
−0.19
2.07
−0.05
−0.06
Appendix
Table 8: Sharpe ratio per stock and its mean over the eight stocks, on 2025-H1 and 2025-FY.
FinMem/FinAgent stocks
Added stocks
Window
Method
TSLA
NFLX
AMZN
MSFT
GS
WMT
BA
XOM
Mean
Buy-and-hold
−48.3
−19.3
−31.2
−20.6
−31.1
−21.9
−26.6
−16.1
−26.9
FinMem, 2-y
−44.3
−21.5
−25.5
−14.0
−31.0
−14.8
0.0
−21.5
−21.6
FinMem, zero
−40.2
−18.3
−17.8
−22.7
−31.0
−13.8
−30.8
−16.1
−23.8
FinAgent, 2-y
−30.6
−20.0
−34.8
−13.3
−20.0
−19.8
−8.6
−23.1
−21.3
FinAgent, zero
−39.1
−14.5
−32.7
−17.7
−22.2
−14.6
−9.0
−15.6
−20.7
Appendix
Table 9: Maximum drawdown (%) per stock and its mean over the eight stocks, on 2025-H1 and 2025-FY; the higher value is the shallower drawdown.
FinMem/FinAgent stocks
Added stocks
Window
Method
TSLA
NFLX
AMZN
MSFT
GS
WMT
BA
XOM
Mean
Buy-and-hold
100
100
100
100
100
100
100
100
100
FinMem, 2-y
61
80
57
47
98
84
0
84
64
FinMem, zero
80
84
45
89
98
64
88
81
79
FinAgent, 2-y
35
69
51
57
65
56
50
59
55
FinAgent, zero
33
68
48
54
65
63
50
66
56
Appendix
Table 10: Exposure per stock, the share (%) of scored days on which the book holds the stock, and its mean over the eight stocks, on 2025-H1 and 2025-FY.
Stock
Window
Manager
Majority
Best seat
best
median
worst
TSLA
2025-H1
5.0
−11.6
Buffett
4.2
−14.4
−26.5
TSLA
2025-FY
14.2
−9.2
Buffett
4.2
−8.1
−28.0
NFLX
2025-H1
54.6
57.0
Livermore
59.4
52.6
37.5
NFLX
2025-FY
16.0
28.0
Livermore
32.8
16.8
−0.8
AMZN
2025-H1
−1.9
−8.5
Dreman
3.2
−3.6
−7.3
AMZN
2025-FY
5.5
−2.7
Dreman
29.4
4.6
−1.4
Appendix
Table 11: Each trader of the default N=6 panel replayed alone, cumulative return (%) without costs, open to open, on 2025-H1 and 2025-FY: the manager’s actions, the majority vote of the panel’s calls, and each trader as if it were the whole panel (BUY = long from the next open, SELL = flat, HOLD = unchanged); best, median and worst are over the six traders.
Stock
Recruited panel ( N=6 )
Fixed
TSLA
Paul Tudor Jones, Jesse Livermore, Jim Simons, George Soros, Nassim Nicholas Taleb, Warren Buffett
5
NFLX
Ed Seykota, Paul Tudor Jones, Jesse Livermore, Jim Simons, William O’Neil, Richard Dennis
4
AMZN
Paul Tudor Jones, David Dreman, Warren Buffett, Jim Simons, George Soros, Jesse Livermore
5
MSFT
Paul Tudor Jones, Linda Bradford Raschke, Jim Simons, Peter Lynch, George Soros, Nassim Nicholas Taleb
3
GS
Jesse Livermore, Paul Tudor Jones, Ed Thorp, Richard Dennis, George Soros, Jim Simons
5
WMT
Warren Buffett, Jim Simons, John Paulson, Richard Dennis, David Shaw, Benjamin Graham
3
Appendix
Table 12: Recruited N=6 panels per stock, and the number of seats shared with the fixed panel.